Anthropic's secret tracker, AI layoffs mount, and model churn accelerates

Anthropic secretly monitored Chinese users, Microsoft cuts 4,800 jobs, Cloudflare adds granular bot controls, and top models now hold their lead for just 7 weeks.

The big picture

Two threads dominate this batch: AI is visibly displacing jobs at scale, and the trust issues around how AI companies handle user data are getting harder to ignore. Neither is theoretical anymore.

Anthropic’s monitoring problem and the broader data trust collapse

Ars Technica reported that Anthropic secretly ran a tracker monitoring Claude users in China, with an engineer describing it as an “experiment” that has since been shut down. (Ars Technica) That framing — “experiment” — is doing a lot of heavy lifting. Covert monitoring of a specific national user group is a significant trust violation, and Anthropic’s brand has been built almost entirely on being the safety-conscious, responsible lab. This undercuts that positioning in a concrete way, not just philosophically.

Meanwhile, Google quietly updated its privacy settings to allow AI training on more user data, and a TechCrunch guide is now walking people through the opt-out process. (TechCrunch) The fact that you need a tutorial to opt out of something this significant says everything about how these settings are designed. If you run services where your users pipe data through Google products, this is worth auditing.

Cloudflare is moving in the opposite direction on transparency. The company replaced its blunt AI bot block with granular controls that let site owners separately manage search crawlers, training crawlers, and agent crawlers. Starting September 15, 2026, training and agent bots will be blocked by default on ad-supported pages. (The Decoder) This is genuinely useful for developers running content sites who want to allow AI search indexing but not wholesale data scraping for model training. Granular is better than blunt.

AI-attributed layoffs are becoming a genre of their own

Microsoft cut roughly 4,800 roles (about 2.1% of its global workforce), with Xbox and commercial sales taking the largest hits. (TechCrunch) TechCrunch is also maintaining a running list of every major 2026 tech layoff where employers explicitly named AI as a contributing factor. (TechCrunch) The pattern is consistent enough that “AI-cited layoff” is now a recognized category of corporate announcement, which is a grim milestone.

The UK’s Financial Conduct Authority is watching this dynamic nervously in financial services, warning of an “arms race” between firms deploying AI and regulators trying to keep pace, while millions of consumers are already making personal finance decisions with AI assistance. (Ars Technica) The FCA is making the case for expanded powers, which is the regulator’s way of saying they feel outgunned. For fintech developers, this is early signal — more prescriptive rules are likely coming.

The model churn problem and who’s fighting for position

Epoch’s capabilities index analysis shows GPT-4 held the top benchmark spot for roughly a year after its 2023 release. Since Claude 3 Opus took the lead in February 2024, the top position has changed hands 17 times, with the median reign lasting just seven weeks. (The Decoder) The headline might suggest accelerating progress, but the fine print is more nuanced: the capability gap between each successive leader is shrinking. Models are cycling faster, but the jumps are getting smaller. For developers, this reinforces Vercel CEO Guillermo Rauch’s point to TechCrunch: when you’re building for production, you care about price/performance, not benchmark rankings, and you want your architecture to swap models without pain. (TechCrunch)

Chinese lab Zhipu AI launched ZCode, a coding environment built on their GLM-5.2 model, positioning it directly against Claude Code and OpenAI Codex at a significantly lower price point. New users get a five-day trial with up to 5 million tokens per day. (The Decoder) Worth a trial if you’re cost-sensitive on agentic coding workflows, though GLM-5.2 is largely untested against the incumbents in real production scenarios.

Quick hits

Apple’s iOS 27 beta adds controls to adjust Siri’s speaking pace and expressivity, part of a broader generative AI rebuild of the assistant. (TechCrunch)
Reddit is using LLMs to detect and remove spam that LLMs are largely responsible for generating — a neat summary of where platform moderation is heading. (TechCrunch)
MIT Technology Review examines Sam Altman’s repeated promises that ordinary Americans will share financially in OpenAI’s success, with the details still conspicuously vague. (MIT Technology Review)
A Verge column on smart glasses privacy uses the Netflix show A Man on the Inside to make a point that Hollywood may have accidentally nailed the real social problem with AI wearables. (The Verge)